Core Concepts
A RESNET50 convolutional neural network can effectively extract optical properties such as reduced scattering coefficient and absorption coefficient from simulated data, with improved accuracy compared to previous approaches using smaller training datasets.
Abstract
The paper presents a study on using a RESNET50 convolutional neural network to extract optical parameters, such as reduced scattering coefficient (μs') and absorption coefficient (μa), from simulated data of light propagation in scattering media.
Key highlights:
The authors used Monte Carlo simulations to generate training data, covering a range of optical properties relevant for biological tissues.
They explored different input representations, including angle and position information of the emerging photons, and found that using both angle and radial position data gave the best results.
Compared to previous works, the RESNET50 network was able to achieve comparable or better reconstruction accuracy using a much smaller training dataset (37,500 samples vs. hundreds of thousands in prior studies).
The authors also trained separate networks to extract scattering and absorption parameters independently, as well as a single network to predict both parameters simultaneously, with no significant difference in performance.
The authors discuss factors that limit the accuracy of the recovered absorption coefficient, such as the loss of photons between the two measurement planes, and suggest ways to potentially improve the approach.
Future work will explore extracting the scattering coefficient and anisotropy factor separately, to gain more insights into the randomization of photons in the scattering medium.
Stats
The sample thickness used in the simulations was 0.118 mm.
The range of μs' was between 0.5 and 2.8 mm-1.
The range of μa was between 0.01 and 1.65 mm-1.
The range of g was between 0.8 and 0.99.
Quotes
"Compared to [11] we apply the image data at an additional plane as well as a network more attuned to extraction of parameters from image data. With this additional plane we aim to supply the network with information on the exit angle of the photon and not simply exit position."
"A very significant aspect of our work is that each dataset only contained 40000 photons compared to 250000 in [11] and 106 in [15], in addition, we used far fewer datasets. We believe that the RESNET architecture is particularly well suited to finding patterns in the data and ignoring noise, so we can extract more useful information for each simulated photon."